Research map / Multi-agent systems and games / LLM agents

RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RADAR, a conditional discrete graph diffusion model that generates multi-agent LLM communication topologies step by step, guided by a redundancy-aware effective-size signal and trained with RL.

Problem

Multi-agent LLM systems use fixed or one-shot-generated communication topologies, which either waste tokens on simple tasks or underperform on complex ones.

Approach

Formulates communication topology design as iterative, node-masking-based discrete graph diffusion conditioned on the query and on an effective-size redundancy measure, with ordering and denoising networks jointly trained via REINFORCE and task-utility policy gradients.

Key results

• RADAR achieves 90.32 average accuracy across six benchmarks with gpt-4o-mini agents, vs 88.57 for ARG-Designer and 85.92 for vanilla single agent.
• On GSM8K RADAR reaches 92.51 and uses about half the tokens of GDesigner (4.2e6 tokens).
• Under prompt attacks RADAR stays nearly unchanged on MMLU (83.14 vs 83.66 clean), while complete graphs drop by up to 4.47%.

Surprising findings

• Several existing multi-agent topologies (e.g., GPTSwarm, ADAS) fail to consistently outperform the single-agent baseline.
• Models trained with a weaker LLM (e.g., gpt-4o-mini) transfer to stronger evaluation LLMs such as DeepSeek-R1, suggesting the generator can be trained cheaply.

Limitations

• Inference time is higher than some baselines due to per-query graph generation (17.55 min vs 7.32 min for AFlow).
• Relies on gpt-4o-mini as the main backbone and only 50 training queries per dataset; utility-based RL needs repeated API evaluation.
• Only a maximum of five agents with fixed role pool is explored; dynamic role generation left to future work.

Experimental setup

Datasets: MMLU GSM8K MultiArith SVAMP AQuA HumanEval

Benchmarks: MMLU GSM8K MultiArith SVAMP AQuA HumanEval

Metrics: accuracy pass@1 token consumption effective size graph density

Baselines: CoT ComplexCoT SC (CoT x5) MultiPersona LLM-Debate LLM-Blender DyLAN AgentVerse MacNet AutoAgents GPTSwarm ADAS

Builds on: GTD ARG-Designer G-Designer AgentPrune DiGress

Closest papers by meaning

• CARD: Towards Conditional Design of Multi-agent Topological Structures (ICLR 2026 · similarity 0.88)
• Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation (AAAI 2026 · similarity 0.88)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.87)
• Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.87)
• AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.87)
• MOC: Multi-Order Communication in LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.87)